Deep Ranking for Style-Aware Room Recommendations (Student Abstract)

Ilkay Yildiz, Esra Ataer Cansizoglu, Hantian Liu, Peter B. Golbus, Ozan Tezcan, Jae-Woo ChoiAAAI 2020 · DOI

Abstract

Identifying interior design style is a highly subjective task, often leading to noisy datasets and poor performance in automated recommendation systems. This paper introduces RoSE v2 (Room Style Estimator), a deep ranking framework that conceptualizes style as a continuous spectrum rather than a set of rigid, discrete categories. By training on relative comparisons between images using a Siamese Network architecture, RoSE v2 mitigates the impact of inter-expert variability and addresses the challenges of data imbalance in niche styles. Experimental results show that RoSE v2 significantly outperforms classification-based baselines, particularly for minority styles such as "Cottage" and "Coastal."

Introduction

In the domain of e-commerce and interior design, "style" is a primary filter for user discovery. However, defining style is inherently subjective; a room labeled as "Modern" by one designer might be categorized as "Traditional" by another. This subjectivity creates two primary hurdles for machine learning:

  1. Inter-expert variability: High noise in training labels due to differing professional opinions.
  2. Data Imbalance: Popular styles (e.g., Modern) dominate datasets, while niche styles (e.g., Coastal or Cottage) suffer from a lack of representative data, causing standard classifiers to overlook them.
An example of a room image from the dataset, illustrating the subjective nature of style classification.

Rather than forcing a single label onto a room, this research proposes a ranking-based approach: instead of asking "What style is this?", the model is trained to answer "Which of these two images is more Modern?"

Methodology

The authors developed RoSE v2, which utilizes a Siamese Network architecture to learn a stylistic embedding space. The core innovation lies in mapping noisy expert labels into comparison labels.

For any pair of images, the model uses the distribution of expert votes to determine a relative ranking. For instance, if Image A received significantly more "Modern" votes than Image B, the network is penalized if it does not rank Image A higher on the Modern spectrum. This approach adapts the Bradley-Terry model—a probabilistic framework for paired comparisons—to a deep learning context. By learning this "style spectrum," the model captures the nuance of rooms that may blend multiple styles (e.g., "mostly Modern with a hint of Traditional").

Dataset and Evaluation

The study utilized a proprietary dataset of room images, each labeled by 10 interior design experts. The dataset focused on four major categories: Modern, Traditional, Cottage, and Coastal. Each image was assigned a label distribution based on the 10 expert votes.

RoSE v2 was compared against two baseline models:

  • RoSE v1n: A standard classifier trained on all noisy multi-expert labels.
  • RoSE v1c: A classifier trained only on "clean" data where experts reached a high level of consensus.

Results and Discussion

RoSE v2 demonstrated superior performance across both majority and minority style classes. Quantitatively, the model achieved a 0.739 recall rate at 1, a 0.792 mAP@5, and a 0.913 recall@5.

Recall rate at 1 across different interior design styles, comparing RoSE v2 against classification baselines.

The most significant gains were observed in minority styles. As shown in the figure above, RoSE v2 improved the recall rate at 1 by 21% for "Cottage" and 15% for "Coastal" relative to the baselines. This suggests that ranking-based losses are more effective at capturing the distinctive features of rare classes than traditional cross-entropy losses, which often bias towards the majority class.

Qualitatively, RoSE v2 proved adept at handling stylistic ambiguity. When presented with a "seed" image where experts disagreed, the model recommended rooms that shared similar levels of ambiguity, effectively mirroring the complex human perception of style.

Conclusion

RoSE v2 represents a shift from discrete classification to continuous ranking in interior design AI. By embracing the subjectivity of style, the framework provides more robust and nuanced recommendations. Future work will focus on scaling the model to include a wider variety of sub-styles and integrating object-level features (e.g., specific furniture pieces) to further refine the style-ranking process.

Summary updated

BibTeX
@inproceedings{yildiz2020deep,
  author = {Ilkay Yildiz and Esra Ataer Cansizoglu and Hantian Liu and Peter B. Golbus and Ozan Tezcan and Jae-Woo Choi},
  title = {{Deep Ranking for Style-Aware Room Recommendations (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i10.7260},
}
Figure